Add comprehensive Vietnamese banking dataset with 3 subsets

- Added 2,471 annotated examples (1,977 train, 494 test)
- Created 3 task-specific subsets:
* Classification: 14 banking aspect categories
* Sentiment: 3-class sentiment analysis
* Aspect-Sentiment: Fine-grained aspect-based sentiment
- Added preprocessing and statistics scripts
- Updated README with detailed dataset documentation
- Added .gitignore for Python files

.gitignore

@@ -0,0 +1,10 @@
+__pycache__/
+*.pyc
+*.pyo
+*.pyd
+.Python
+.venv/
+venv/
+ENV/
+env/
+.DS_Store

.python-version

@@ -0,0 +1 @@
+3.13

README.md

@@ -9,17 +9,32 @@ tags:
 - banking
 - finance
 - nlp
+- sentiment-analysis
+- aspect-based-sentiment-analysis
 task_categories:
 - text-classification
+- sentiment-classification
 size_categories:
-- n<1K
+- 1K<n<10K
 configs:
-- config_name: default
+- config_name: classification
   data_files:
   - split: train
-    path: data/train.txt
+    path: data/classification/train.jsonl
   - split: test
-    path: data/test.txt
+    path: data/classification/test.jsonl
+- config_name: sentiment
+  data_files:
+  - split: train
+    path: data/sentiment/train.jsonl
+  - split: test
+    path: data/sentiment/test.jsonl
+- config_name: aspect_sentiment
+  data_files:
+  - split: train
+    path: data/aspect_sentiment/train.jsonl
+  - split: test
+    path: data/aspect_sentiment/test.jsonl
 ---
 
 # UTS2017_Bank Dataset
@@ -28,14 +43,23 @@ configs:
 
 ### Dataset Summary
 
-The UTS2017_Bank dataset is a Vietnamese language dataset specifically focused on banking and financial domain text. It was created as part of the UnderTheSea NLP project to support Vietnamese natural language processing tasks in the financial sector. The dataset provides domain-specific text samples that can be used for various NLP tasks including text generation, classification, and named entity recognition in the banking context.
+The UTS2017_Bank dataset is a comprehensive Vietnamese banking domain dataset containing customer feedback and reviews about banking services. It contains **2,471 annotated examples** (1,977 train, 494 test) with both aspect labels and sentiment annotations. The dataset supports multiple NLP tasks including aspect classification, sentiment analysis, and aspect-based sentiment analysis in the Vietnamese banking sector.
+
+### Supported Tasks
+
+The dataset is provided in three subsets for different tasks:
+
+1. **Classification** (`classification`): Banking aspect classification
+   - 14 aspect categories (CUSTOMER_SUPPORT, TRADEMARK, LOAN, etc.)
+   - Train: 1,977 examples | Test: 494 examples
 
-### Supported Tasks and Leaderboards
+2. **Sentiment Analysis** (`sentiment`): Overall sentiment classification
+   - 3 sentiment classes: positive (61.3%), negative (37.6%), neutral (1.2%)
+   - Train: 1,977 examples | Test: 494 examples
 
-- **Text Generation**: Generate banking-related content in Vietnamese
-- **Text Classification**: Classify banking documents and queries
-- **Named Entity Recognition**: Identify financial entities, banking terms, and monetary values
-- **Language Modeling**: Train Vietnamese language models specialized for banking domain
+3. **Aspect-Based Sentiment** (`aspect_sentiment`): Fine-grained aspect-sentiment pairs
+   - Multi-aspect support (1.5% examples have multiple aspects)
+   - Train: 1,977 examples | Test: 494 examples
 
 ### Languages
 
@@ -45,83 +69,170 @@ The dataset is exclusively in Vietnamese (`vi`).
 
 ### Data Instances
 
-Each instance in the dataset contains a single text field with Vietnamese banking-related content. Example:
+Each subset has a different structure:
+
+**Classification subset:**
+```json
+{
+  "text": "Hotline khó gọi quá gọi mãi ko thưa máy à",
+  "label": "CUSTOMER_SUPPORT"
+}
+```
+
+**Sentiment subset:**
+```json
+{
+  "text": "Dịch vụ tiện dụng quá!",
+  "sentiment": "positive"
+}
+```
 
+**Aspect-Sentiment subset:**
 ```json
 {
-  "text": "Ngân hàng Nhà nước Việt Nam"
+  "text": "Mình xài cái thể VISA của BIDV hạn mức 100tr...",
+  "aspects": [
+    {"aspect": "CARD", "sentiment": "negative"},
+    {"aspect": "CUSTOMER_SUPPORT", "sentiment": "negative"}
+  ]
 }
 ```
 
 ### Data Fields
 
-- `text` (string): A Vietnamese text string containing banking or finance-related content
+- **Classification subset:**
+  - `text` (string): Customer feedback text in Vietnamese
+  - `label` (string): Banking aspect category
+
+- **Sentiment subset:**
+  - `text` (string): Customer feedback text in Vietnamese
+  - `sentiment` (string): Overall sentiment (positive/negative/neutral)
+
+- **Aspect-Sentiment subset:**
+  - `text` (string): Customer feedback text in Vietnamese
+  - `aspects` (list): List of aspect-sentiment pairs
+
+### Aspect Categories
+
+The dataset contains 14 banking aspect categories:
+
+| Aspect | Train Count | Test Count | Description |
+|--------|------------|------------|-------------|
+| CUSTOMER_SUPPORT | 774 (39.2%) | 338 (68.4%) | Customer service quality |
+| TRADEMARK | 699 (35.4%) | 41 (8.3%) | Bank brand and reputation |
+| LOAN | 74 (3.7%) | 3 (0.6%) | Loan services |
+| INTERNET_BANKING | 70 (3.5%) | 32 (6.5%) | Online banking services |
+| CARD | 66 (3.3%) | 44 (8.9%) | Credit/debit card services |
+| INTEREST_RATE | 60 (3.0%) | 6 (1.2%) | Interest rates |
+| PROMOTION | 53 (2.7%) | 9 (1.8%) | Promotional offers |
+| OTHER | 69 (3.5%) | 12 (2.4%) | Other topics |
+| DISCOUNT | 41 (2.1%) | 2 (0.4%) | Discounts and benefits |
+| MONEY_TRANSFER | 34 (1.7%) | 2 (0.4%) | Money transfer services |
+| PAYMENT | 15 (0.8%) | 2 (0.4%) | Payment services |
+| SAVING | 13 (0.7%) | 3 (0.6%) | Savings accounts |
+| ACCOUNT | 5 (0.3%) | 0 (0%) | Account management |
+| SECURITY | 4 (0.2%) | 0 (0%) | Security concerns |
+
+### Data Statistics
+
+- **Text Length:**
+  - Train: Average 23.8 words (median 14, range 1-816)
+  - Test: Average 30.6 words (median 18, range 1-411)
+
+- **Sentiment Distribution:**
+  - Train: Positive 61.3%, Negative 37.6%, Neutral 1.2%
+  - Test: Negative 60.9%, Positive 37.4%, Neutral 1.6%
 
-### Data Splits
+## Usage
 
-The dataset is divided into two splits:
+### Loading the Dataset
 
-| Split | Number of Examples |
-|-------|-------------------|
-| train | 2 |
-| test  | 2 |
+```python
+from datasets import load_dataset
 
-*Note: This is a sample dataset. The full version may contain more examples.*
+# Load classification subset
+dataset_clf = load_dataset("undertheseanlp/UTS2017_Bank", "classification")
+
+# Load sentiment subset
+dataset_sent = load_dataset("undertheseanlp/UTS2017_Bank", "sentiment")
+
+# Load aspect-sentiment subset
+dataset_aspect = load_dataset("undertheseanlp/UTS2017_Bank", "aspect_sentiment")
+```
+
+### Data Preprocessing
+
+A preprocessing script is available in the repository:
+
+```python
+from preprocess_data import process_banking_data
+
+# Process raw data into three subsets
+process_banking_data("raw_data/train.txt", "data")
+```
 
 ## Dataset Creation
 
 ### Curation Rationale
 
-This dataset was created to address the lack of domain-specific Vietnamese NLP resources in the banking and financial sector. It aims to support the development of Vietnamese language models and applications specifically tailored for banking services, financial advisory, and customer service automation in Vietnam's banking industry.
+This dataset addresses the lack of Vietnamese NLP resources in the banking sector, supporting the development of specialized models for:
+- Customer feedback analysis
+- Service quality monitoring
+- Banking chatbots and virtual assistants
+- Financial sentiment analysis
 
 ### Source Data
 
-#### Initial Data Collection and Normalization
-
-The text samples were collected from public banking communications, financial documents, and banking service descriptions in Vietnamese. All texts have been normalized for consistency in formatting and encoding.
+The dataset contains real customer feedback and reviews about Vietnamese banking services, collected from public sources and banking communications. All texts have been anonymized to remove personal information.
 
 ### Annotations
 
-This dataset contains raw text without additional annotations. Users can apply their own annotation schemes based on their specific use cases.
+The dataset includes:
+- 14 banking-specific aspect categories
+- 3-class sentiment labels (positive, negative, neutral)
+- Support for multi-aspect annotations (1.5% of examples)
 
 ### Personal and Sensitive Information
 
-The dataset has been carefully reviewed to ensure no personal customer information, account numbers, or sensitive financial data is included. All text samples contain only general banking terminology and public information.
+All personal information, account numbers, and sensitive financial data have been removed. The dataset contains only general banking terminology and anonymized feedback.
 
 ## Considerations for Using the Data
 
-### Social Impact of Dataset
+### Social Impact
 
-This dataset can contribute to:
-- Improving Vietnamese banking services through better NLP tools
-- Enhancing financial inclusion by enabling better Vietnamese language support in digital banking
-- Supporting research in domain-specific Vietnamese NLP
+This dataset enables:
+- Better understanding of customer needs in Vietnamese banking
+- Improved customer service through automated analysis
+- Enhanced financial inclusion through better language support
 
 ### Discussion of Biases
 
-The dataset may reflect common patterns and terminology used in formal Vietnamese banking communications, which might not fully represent colloquial or regional variations in how people discuss banking topics.
+- **Language style**: Primarily informal customer feedback language
+- **Sentiment imbalance**: More positive examples in training set
+- **Aspect distribution**: Heavy skew towards CUSTOMER_SUPPORT and TRADEMARK
 
-### Other Known Limitations
+### Known Limitations
 
-- Limited size: This is a sample dataset with a small number of examples
-- Domain specificity: The dataset is highly specialized for banking and may not generalize well to other domains
-- Formal language bias: The text samples primarily use formal banking Vietnamese
+- Limited neutral sentiment examples (1-2%)
+- Aspect distribution varies between train and test sets
+- Multi-aspect examples are rare (1.5%)
+- Domain-specific to banking sector
 
 ## Additional Information
 
 ### Dataset Curators
 
-This dataset was created and is maintained by the UnderTheSea NLP team, a community project focused on developing Vietnamese NLP resources and tools.
+Created and maintained by the UnderTheSea NLP team, focusing on Vietnamese NLP resources and tools development.
 
 ### Licensing Information
 
-This dataset is released under the Apache 2.0 License.
+Released under the Apache 2.0 License.
 
 ### Citation Information
 
 ```bibtex
 @dataset{uts2017_bank,
-  title={UTS2017_Bank: A Vietnamese Banking Domain Dataset},
+  title={UTS2017_Bank: A Vietnamese Banking Domain Dataset for Aspect-Based Sentiment Analysis},
   author={UnderTheSea NLP},
   year={2017},
   publisher={Hugging Face},
@@ -131,13 +242,14 @@ This dataset is released under the Apache 2.0 License.
 
 ### Contributions
 
-Thanks to the UnderTheSea NLP community for creating and maintaining this dataset. Special thanks to all contributors who helped collect, clean, and validate the banking domain texts.
+Thanks to the UnderTheSea NLP community for creating and maintaining this dataset.
 
 ## Contact
 
-For questions, issues, or contributions, please:
+For questions or contributions:
 - Open an issue on the [dataset repository](https://huggingface.co/datasets/undertheseanlp/UTS2017_Bank/discussions)
+- Visit [UnderTheSea NLP](https://github.com/undertheseanlp)
 
 ## Updates and Versions
 
-- **Version 1.0.0** (Current): Initial release with sample banking texts
+- **Version 1.0.0** (Current): Initial release with 2,471 annotated banking feedback examples in three task-specific subsets

data/test.txt

@@ -1,2 +0,0 @@
-Chuyển khoản nhanh 24/7
-Vay tín chấp không thế chấp

data/train.txt

@@ -1,2 +0,0 @@
-Ngân hàng Nhà nước Việt Nam
-Tài khoản tiết kiệm lãi suất cao

dataset_statistics.py

@@ -0,0 +1,227 @@
+import json
+from pathlib import Path
+from collections import Counter
+import statistics as stats
+
+
+def load_jsonl(file_path):
+    """Load JSONL file and return list of items."""
+    items = []
+    with open(file_path, 'r', encoding='utf-8') as f:
+        for line in f:
+            items.append(json.loads(line.strip()))
+    return items
+
+
+def calculate_text_statistics(items):
+    """Calculate statistics for text fields."""
+    text_lengths = [len(item['text'].split()) for item in items]
+    char_lengths = [len(item['text']) for item in items]
+
+    return {
+        'avg_words': stats.mean(text_lengths),
+        'min_words': min(text_lengths),
+        'max_words': max(text_lengths),
+        'median_words': stats.median(text_lengths),
+        'avg_chars': stats.mean(char_lengths),
+        'min_chars': min(char_lengths),
+        'max_chars': max(char_lengths),
+        'median_chars': stats.median(char_lengths),
+    }
+
+
+def analyze_classification_subset():
+    """Analyze classification subset statistics."""
+    print("\n" + "="*60)
+    print("CLASSIFICATION SUBSET ANALYSIS")
+    print("="*60)
+
+    for split in ['train', 'test']:
+        file_path = Path(f'data/classification/{split}.jsonl')
+        items = load_jsonl(file_path)
+
+        print(f"\n{split.upper()} Split:")
+        print(f"  Total examples: {len(items)}")
+
+        # Label distribution
+        label_counter = Counter(item['label'] for item in items)
+        print("\n  Label Distribution:")
+        for label, count in label_counter.most_common():
+            percentage = (count / len(items)) * 100
+            print(f"    {label:20s}: {count:4d} ({percentage:5.1f}%)")
+
+        # Text statistics
+        text_stats = calculate_text_statistics(items)
+        print("\n  Text Statistics:")
+        print(f"    Words per text - Avg: {text_stats['avg_words']:.1f}, "
+              f"Min: {text_stats['min_words']}, Max: {text_stats['max_words']}, "
+              f"Median: {text_stats['median_words']:.1f}")
+        print(f"    Chars per text - Avg: {text_stats['avg_chars']:.1f}, "
+              f"Min: {text_stats['min_chars']}, Max: {text_stats['max_chars']}, "
+              f"Median: {text_stats['median_chars']:.1f}")
+
+
+def analyze_sentiment_subset():
+    """Analyze sentiment subset statistics."""
+    print("\n" + "="*60)
+    print("SENTIMENT SUBSET ANALYSIS")
+    print("="*60)
+
+    for split in ['train', 'test']:
+        file_path = Path(f'data/sentiment/{split}.jsonl')
+        items = load_jsonl(file_path)
+
+        print(f"\n{split.upper()} Split:")
+        print(f"  Total examples: {len(items)}")
+
+        # Sentiment distribution
+        sentiment_counter = Counter(item['sentiment'] for item in items)
+        print("\n  Sentiment Distribution:")
+        for sentiment, count in sentiment_counter.most_common():
+            percentage = (count / len(items)) * 100
+            print(f"    {sentiment:10s}: {count:4d} ({percentage:5.1f}%)")
+
+        # Text statistics
+        text_stats = calculate_text_statistics(items)
+        print("\n  Text Statistics:")
+        print(f"    Words per text - Avg: {text_stats['avg_words']:.1f}, "
+              f"Min: {text_stats['min_words']}, Max: {text_stats['max_words']}, "
+              f"Median: {text_stats['median_words']:.1f}")
+
+
+def analyze_aspect_sentiment_subset():
+    """Analyze aspect-sentiment subset statistics."""
+    print("\n" + "="*60)
+    print("ASPECT-SENTIMENT SUBSET ANALYSIS")
+    print("="*60)
+
+    for split in ['train', 'test']:
+        file_path = Path(f'data/aspect_sentiment/{split}.jsonl')
+        items = load_jsonl(file_path)
+
+        print(f"\n{split.upper()} Split:")
+        print(f"  Total examples: {len(items)}")
+
+        # Multi-aspect analysis
+        single_aspect = sum(1 for item in items if len(item['aspects']) == 1)
+        multi_aspect = sum(1 for item in items if len(item['aspects']) > 1)
+        max_aspects = max(len(item['aspects']) for item in items)
+
+        print(f"\n  Aspect Coverage:")
+        print(f"    Single aspect: {single_aspect} ({(single_aspect/len(items))*100:.1f}%)")
+        print(f"    Multi-aspect: {multi_aspect} ({(multi_aspect/len(items))*100:.1f}%)")
+        print(f"    Max aspects per example: {max_aspects}")
+
+        # Aspect-sentiment pair distribution
+        aspect_sentiment_pairs = []
+        for item in items:
+            for asp in item['aspects']:
+                aspect_sentiment_pairs.append(f"{asp['aspect']}#{asp['sentiment']}")
+
+        pair_counter = Counter(aspect_sentiment_pairs)
+        print("\n  Top 10 Aspect-Sentiment Pairs:")
+        for pair, count in pair_counter.most_common(10):
+            aspect, sentiment = pair.split('#')
+            percentage = (count / len(aspect_sentiment_pairs)) * 100
+            print(f"    {aspect:20s} + {sentiment:8s}: {count:4d} ({percentage:5.1f}%)")
+
+        # Overall aspect distribution
+        aspect_counter = Counter()
+        sentiment_by_aspect = {}
+
+        for item in items:
+            for asp in item['aspects']:
+                aspect = asp['aspect']
+                sentiment = asp['sentiment']
+                aspect_counter[aspect] += 1
+
+                if aspect not in sentiment_by_aspect:
+                    sentiment_by_aspect[aspect] = Counter()
+                sentiment_by_aspect[aspect][sentiment] += 1
+
+        print("\n  Aspect Distribution with Sentiment Breakdown:")
+        for aspect, count in aspect_counter.most_common():
+            percentage = (count / sum(aspect_counter.values())) * 100
+            print(f"\n    {aspect:20s}: {count:4d} ({percentage:5.1f}%)")
+
+            # Sentiment breakdown for this aspect
+            sentiments = sentiment_by_aspect[aspect]
+            total_aspect = sum(sentiments.values())
+            for sentiment in ['positive', 'negative', 'neutral']:
+                if sentiment in sentiments:
+                    sent_count = sentiments[sentiment]
+                    sent_pct = (sent_count / total_aspect) * 100
+                    print(f"      - {sentiment:8s}: {sent_count:3d} ({sent_pct:5.1f}%)")
+
+
+def generate_summary_statistics():
+    """Generate overall summary statistics."""
+    print("\n" + "="*60)
+    print("DATASET SUMMARY")
+    print("="*60)
+
+    total_train = len(load_jsonl('data/classification/train.jsonl'))
+    total_test = len(load_jsonl('data/classification/test.jsonl'))
+
+    print("\nTotal Dataset Size:")
+    print(f"  Train: {total_train} examples")
+    print(f"  Test: {total_test} examples")
+    print(f"  Total: {total_train + total_test} examples")
+    print(f"  Train/Test Ratio: {total_train/total_test:.2f}:1")
+
+    # Available subsets
+    print("\nAvailable Subsets:")
+    print("  1. Classification: Text → Label (14 banking aspect categories)")
+    print("  2. Sentiment: Text → Sentiment (positive/negative/neutral)")
+    print("  3. Aspect-Sentiment: Text → Multiple (Aspect, Sentiment) pairs")
+
+    # Data format
+    print("\nData Format:")
+    print("  - All subsets use JSONL format")
+    print("  - UTF-8 encoding")
+    print("  - Vietnamese language text")
+
+    # Use cases
+    print("\nRecommended Use Cases:")
+    print("  - Classification: Banking domain text classification")
+    print("  - Sentiment: Customer feedback sentiment analysis")
+    print("  - Aspect-Sentiment: Fine-grained aspect-based sentiment analysis")
+
+
+def save_statistics_report():
+    """Save statistics to a markdown file."""
+    import sys
+    from io import StringIO
+
+    # Capture output
+    old_stdout = sys.stdout
+    sys.stdout = buffer = StringIO()
+
+    # Run all analyses
+    generate_summary_statistics()
+    analyze_classification_subset()
+    analyze_sentiment_subset()
+    analyze_aspect_sentiment_subset()
+
+    # Get output
+    output = buffer.getvalue()
+    sys.stdout = old_stdout
+
+    # Save to file
+    with open('statistics_report.md', 'w', encoding='utf-8') as f:
+        f.write("# UTS2017_Bank Dataset Statistics Report\n\n")
+        f.write("```\n")
+        f.write(output)
+        f.write("```\n")
+
+    print("Statistics report saved to statistics_report.md")
+
+
+if __name__ == "__main__":
+    generate_summary_statistics()
+    analyze_classification_subset()
+    analyze_sentiment_subset()
+    analyze_aspect_sentiment_subset()
+
+    print("\n" + "="*60)
+    save_statistics_report()

preprocess_data.py

@@ -0,0 +1,143 @@
+import re
+import json
+from pathlib import Path
+
+
+def process_banking_data(input_file, output_dir):
+    """
+    Process banking data into three subsets:
+    1. Classification: text -> label (main aspect only)
+    2. Sentiment: text -> sentiment (overall sentiment)
+    3. Aspect-Sentiment: text -> aspect, sentiment pairs
+    """
+
+    classification_data = []
+    sentiment_data = []
+    aspect_sentiment_data = []
+
+    with open(input_file, 'r', encoding='utf-8') as f:
+        for line_num, line in enumerate(f, 1):
+            line = line.strip()
+            if not line:
+                continue
+
+            # Extract labels and sentiments
+            label_pattern = r'__label__([A-Z_]+)#(positive|negative|neutral)'
+            matches = re.findall(label_pattern, line)
+
+            # Remove labels from text
+            text = re.sub(r'__label__[A-Z_]+#(positive|negative|neutral)\s*', '', line).strip()
+
+            if not text or not matches:
+                print(f"Skipping line {line_num}: No valid data found")
+                continue
+
+            # Extract aspects and sentiments
+            aspects = [m[0] for m in matches]
+            sentiments = [m[1] for m in matches]
+
+            # 1. Classification subset (first aspect as main label)
+            classification_data.append({
+                "text": text,
+                "label": aspects[0]
+            })
+
+            # 2. Sentiment-only subset (overall sentiment)
+            # If multiple sentiments, use the first one or most frequent
+            overall_sentiment = sentiments[0]
+            if len(set(sentiments)) == 1:
+                overall_sentiment = sentiments[0]
+            else:
+                # Use most common sentiment
+                sentiment_counts = {}
+                for s in sentiments:
+                    sentiment_counts[s] = sentiment_counts.get(s, 0) + 1
+                overall_sentiment = max(sentiment_counts, key=sentiment_counts.get)
+
+            sentiment_data.append({
+                "text": text,
+                "sentiment": overall_sentiment
+            })
+
+            # 3. Aspect-Sentiment subset
+            aspect_sentiment_pairs = []
+            for aspect, sentiment in zip(aspects, sentiments):
+                aspect_sentiment_pairs.append({
+                    "aspect": aspect,
+                    "sentiment": sentiment
+                })
+
+            aspect_sentiment_data.append({
+                "text": text,
+                "aspects": aspect_sentiment_pairs
+            })
+
+    # Save the three subsets
+    output_dir = Path(output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    # Determine if this is train or test based on input filename
+    split = "train" if "train" in str(input_file) else "test"
+
+    # Save classification subset
+    classification_file = output_dir / "classification" / f"{split}.jsonl"
+    classification_file.parent.mkdir(parents=True, exist_ok=True)
+    with open(classification_file, 'w', encoding='utf-8') as f:
+        for item in classification_data:
+            f.write(json.dumps(item, ensure_ascii=False) + '\n')
+    print(f"Saved {len(classification_data)} classification examples to {classification_file}")
+
+    # Save sentiment subset
+    sentiment_file = output_dir / "sentiment" / f"{split}.jsonl"
+    sentiment_file.parent.mkdir(parents=True, exist_ok=True)
+    with open(sentiment_file, 'w', encoding='utf-8') as f:
+        for item in sentiment_data:
+            f.write(json.dumps(item, ensure_ascii=False) + '\n')
+    print(f"Saved {len(sentiment_data)} sentiment examples to {sentiment_file}")
+
+    # Save aspect-sentiment subset
+    aspect_sentiment_file = output_dir / "aspect_sentiment" / f"{split}.jsonl"
+    aspect_sentiment_file.parent.mkdir(parents=True, exist_ok=True)
+    with open(aspect_sentiment_file, 'w', encoding='utf-8') as f:
+        for item in aspect_sentiment_data:
+            f.write(json.dumps(item, ensure_ascii=False) + '\n')
+    print(f"Saved {len(aspect_sentiment_data)} aspect-sentiment examples to {aspect_sentiment_file}")
+
+    # Print statistics
+    print("\n=== Statistics ===")
+    print(f"Total examples processed: {len(classification_data)}")
+
+    # Label distribution
+    label_counts = {}
+    for item in classification_data:
+        label = item['label']
+        label_counts[label] = label_counts.get(label, 0) + 1
+
+    print("\nLabel distribution:")
+    for label, count in sorted(label_counts.items(), key=lambda x: x[1], reverse=True):
+        print(f"  {label}: {count}")
+
+    # Sentiment distribution
+    sentiment_counts = {}
+    for item in sentiment_data:
+        sentiment = item['sentiment']
+        sentiment_counts[sentiment] = sentiment_counts.get(sentiment, 0) + 1
+
+    print("\nSentiment distribution:")
+    for sentiment, count in sorted(sentiment_counts.items(), key=lambda x: x[1], reverse=True):
+        print(f"  {sentiment}: {count}")
+
+    # Multi-aspect examples
+    multi_aspect_count = sum(1 for item in aspect_sentiment_data if len(item['aspects']) > 1)
+    print(f"\nExamples with multiple aspects: {multi_aspect_count}")
+
+
+if __name__ == "__main__":
+    # Process train data
+    print("Processing train data...")
+    process_banking_data("raw_data/train.txt", "data")
+
+    # Process test data
+    print("\n" + "="*50)
+    print("Processing test data...")
+    process_banking_data("raw_data/test.txt", "data")

pyproject.toml

@@ -0,0 +1,8 @@
+[project]
+name = "uts2017-bank"
+version = "1.0.0"
+description = "Underthesea 2017 Bank Dataset"
+requires-python = ">=3.13"
+dependencies = [
+    "datasets>=4.1.1",
+]

statistics_report.md

@@ -0,0 +1,262 @@
+# UTS2017_Bank Dataset Statistics Report
+
+```
+
+============================================================
+DATASET SUMMARY
+============================================================
+
+Total Dataset Size:
+  Train: 1977 examples
+  Test: 494 examples
+  Total: 2471 examples
+  Train/Test Ratio: 4.00:1
+
+Available Subsets:
+  1. Classification: Text → Label (14 banking aspect categories)
+  2. Sentiment: Text → Sentiment (positive/negative/neutral)
+  3. Aspect-Sentiment: Text → Multiple (Aspect, Sentiment) pairs
+
+Data Format:
+  - All subsets use JSONL format
+  - UTF-8 encoding
+  - Vietnamese language text
+
+Recommended Use Cases:
+  - Classification: Banking domain text classification
+  - Sentiment: Customer feedback sentiment analysis
+  - Aspect-Sentiment: Fine-grained aspect-based sentiment analysis
+
+============================================================
+CLASSIFICATION SUBSET ANALYSIS
+============================================================
+
+TRAIN Split:
+  Total examples: 1977
+
+  Label Distribution:
+    CUSTOMER_SUPPORT    :  774 ( 39.2%)
+    TRADEMARK           :  699 ( 35.4%)
+    LOAN                :   74 (  3.7%)
+    INTERNET_BANKING    :   70 (  3.5%)
+    OTHER               :   69 (  3.5%)
+    CARD                :   66 (  3.3%)
+    INTEREST_RATE       :   60 (  3.0%)
+    PROMOTION           :   53 (  2.7%)
+    DISCOUNT            :   41 (  2.1%)
+    MONEY_TRANSFER      :   34 (  1.7%)
+    PAYMENT             :   15 (  0.8%)
+    SAVING              :   13 (  0.7%)
+    ACCOUNT             :    5 (  0.3%)
+    SECURITY            :    4 (  0.2%)
+
+  Text Statistics:
+    Words per text - Avg: 23.8, Min: 1, Max: 816, Median: 14.0
+    Chars per text - Avg: 106.3, Min: 3, Max: 3787, Median: 62.0
+
+TEST Split:
+  Total examples: 494
+
+  Label Distribution:
+    CUSTOMER_SUPPORT    :  338 ( 68.4%)
+    CARD                :   44 (  8.9%)
+    TRADEMARK           :   41 (  8.3%)
+    INTERNET_BANKING    :   32 (  6.5%)
+    OTHER               :   12 (  2.4%)
+    PROMOTION           :    9 (  1.8%)
+    INTEREST_RATE       :    6 (  1.2%)
+    LOAN                :    3 (  0.6%)
+    SAVING              :    3 (  0.6%)
+    MONEY_TRANSFER      :    2 (  0.4%)
+    DISCOUNT            :    2 (  0.4%)
+    PAYMENT             :    2 (  0.4%)
+
+  Text Statistics:
+    Words per text - Avg: 30.6, Min: 1, Max: 411, Median: 18.0
+    Chars per text - Avg: 134.8, Min: 4, Max: 1800, Median: 80.0
+
+============================================================
+SENTIMENT SUBSET ANALYSIS
+============================================================
+
+TRAIN Split:
+  Total examples: 1977
+
+  Sentiment Distribution:
+    positive  : 1211 ( 61.3%)
+    negative  :  743 ( 37.6%)
+    neutral   :   23 (  1.2%)
+
+  Text Statistics:
+    Words per text - Avg: 23.8, Min: 1, Max: 816, Median: 14.0
+
+TEST Split:
+  Total examples: 494
+
+  Sentiment Distribution:
+    negative  :  301 ( 60.9%)
+    positive  :  185 ( 37.4%)
+    neutral   :    8 (  1.6%)
+
+  Text Statistics:
+    Words per text - Avg: 30.6, Min: 1, Max: 411, Median: 18.0
+
+============================================================
+ASPECT-SENTIMENT SUBSET ANALYSIS
+============================================================
+
+TRAIN Split:
+  Total examples: 1977
+
+  Aspect Coverage:
+    Single aspect: 1948 (98.5%)
+    Multi-aspect: 29 (1.5%)
+    Max aspects per example: 4
+
+  Top 10 Aspect-Sentiment Pairs:
+    TRADEMARK            + positive:  652 ( 32.5%)
+    CUSTOMER_SUPPORT     + positive:  409 ( 20.4%)
+    CUSTOMER_SUPPORT     + negative:  370 ( 18.4%)
+    INTEREST_RATE        + negative:   63 (  3.1%)
+    LOAN                 + negative:   61 (  3.0%)
+    INTERNET_BANKING     + negative:   57 (  2.8%)
+    CARD                 + negative:   54 (  2.7%)
+    TRADEMARK            + negative:   47 (  2.3%)
+    OTHER                + negative:   35 (  1.7%)
+    PROMOTION            + positive:   33 (  1.6%)
+
+  Aspect Distribution with Sentiment Breakdown:
+
+    CUSTOMER_SUPPORT    :  784 ( 39.0%)
+      - positive: 409 ( 52.2%)
+      - negative: 370 ( 47.2%)
+      - neutral :   5 (  0.6%)
+
+    TRADEMARK           :  699 ( 34.8%)
+      - positive: 652 ( 93.3%)
+      - negative:  47 (  6.7%)
+
+    INTERNET_BANKING    :   79 (  3.9%)
+      - positive:  20 ( 25.3%)
+      - negative:  57 ( 72.2%)
+      - neutral :   2 (  2.5%)
+
+    LOAN                :   74 (  3.7%)
+      - positive:  13 ( 17.6%)
+      - negative:  61 ( 82.4%)
+
+    OTHER               :   69 (  3.4%)
+      - positive:  30 ( 43.5%)
+      - negative:  35 ( 50.7%)
+      - neutral :   4 (  5.8%)
+
+    INTEREST_RATE       :   68 (  3.4%)
+      - positive:   4 (  5.9%)
+      - negative:  63 ( 92.6%)
+      - neutral :   1 (  1.5%)
+
+    CARD                :   67 (  3.3%)
+      - positive:  12 ( 17.9%)
+      - negative:  54 ( 80.6%)
+      - neutral :   1 (  1.5%)
+
+    PROMOTION           :   53 (  2.6%)
+      - positive:  33 ( 62.3%)
+      - negative:  17 ( 32.1%)
+      - neutral :   3 (  5.7%)
+
+    DISCOUNT            :   42 (  2.1%)
+      - positive:  20 ( 47.6%)
+      - negative:  18 ( 42.9%)
+      - neutral :   4 (  9.5%)
+
+    MONEY_TRANSFER      :   36 (  1.8%)
+      - positive:   5 ( 13.9%)
+      - negative:  31 ( 86.1%)
+
+    PAYMENT             :   15 (  0.7%)
+      - positive:  11 ( 73.3%)
+      - negative:   4 ( 26.7%)
+
+    SAVING              :   13 (  0.6%)
+      - positive:   6 ( 46.2%)
+      - negative:   6 ( 46.2%)
+      - neutral :   1 (  7.7%)
+
+    ACCOUNT             :    5 (  0.2%)
+      - negative:   5 (100.0%)
+
+    SECURITY            :    5 (  0.2%)
+      - positive:   1 ( 20.0%)
+      - negative:   2 ( 40.0%)
+      - neutral :   2 ( 40.0%)
+
+TEST Split:
+  Total examples: 494
+
+  Aspect Coverage:
+    Single aspect: 492 (99.6%)
+    Multi-aspect: 2 (0.4%)
+    Max aspects per example: 3
+
+  Top 10 Aspect-Sentiment Pairs:
+    CUSTOMER_SUPPORT     + negative:  215 ( 43.3%)
+    CUSTOMER_SUPPORT     + positive:  122 ( 24.5%)
+    CARD                 + negative:   37 (  7.4%)
+    TRADEMARK            + positive:   35 (  7.0%)
+    INTERNET_BANKING     + negative:   27 (  5.4%)
+    PROMOTION            + positive:    7 (  1.4%)
+    TRADEMARK            + negative:    6 (  1.2%)
+    OTHER                + positive:    6 (  1.2%)
+    INTERNET_BANKING     + positive:    6 (  1.2%)
+    INTEREST_RATE        + negative:    6 (  1.2%)
+
+  Aspect Distribution with Sentiment Breakdown:
+
+    CUSTOMER_SUPPORT    :  339 ( 68.2%)
+      - positive: 122 ( 36.0%)
+      - negative: 215 ( 63.4%)
+      - neutral :   2 (  0.6%)
+
+    CARD                :   44 (  8.9%)
+      - positive:   5 ( 11.4%)
+      - negative:  37 ( 84.1%)
+      - neutral :   2 (  4.5%)
+
+    TRADEMARK           :   41 (  8.2%)
+      - positive:  35 ( 85.4%)
+      - negative:   6 ( 14.6%)
+
+    INTERNET_BANKING    :   33 (  6.6%)
+      - positive:   6 ( 18.2%)
+      - negative:  27 ( 81.8%)
+
+    OTHER               :   12 (  2.4%)
+      - positive:   6 ( 50.0%)
+      - negative:   3 ( 25.0%)
+      - neutral :   3 ( 25.0%)
+
+    PROMOTION           :    9 (  1.8%)
+      - positive:   7 ( 77.8%)
+      - negative:   1 ( 11.1%)
+      - neutral :   1 ( 11.1%)
+
+    INTEREST_RATE       :    6 (  1.2%)
+      - negative:   6 (100.0%)
+
+    LOAN                :    3 (  0.6%)
+      - negative:   3 (100.0%)
+
+    MONEY_TRANSFER      :    3 (  0.6%)
+      - negative:   3 (100.0%)
+
+    SAVING              :    3 (  0.6%)
+      - positive:   3 (100.0%)
+
+    DISCOUNT            :    2 (  0.4%)
+      - negative:   2 (100.0%)
+
+    PAYMENT             :    2 (  0.4%)
+      - positive:   1 ( 50.0%)
+      - negative:   1 ( 50.0%)
+```